Harness Hub
All people

Researcher

Joel Niklaus

AI researcher

Evals & benchmarks

Researcher who ran ten harnesses on SWE-bench Pro and showed that swapping the harness moved pass@1 from 23% to 52% on the same model, a key data point for "harness matters".

X

Sources

  1. x.com/joelniklaus/status/2085725862142623875

Verified Oct 29, 2026